rajatagarwal457/pg320-forecast-eval

pg-320 — Politics / Geopolitics Forecasting Eval

A 320-question yes/no forecasting benchmark used to evaluate the Anthral Research GRPO LoRA adapters against frontier closed-source baselines.

Each question is bundled with the exact retrieval context that was used during the published evaluation runs — top-N news chunks, all dated strictly before the question creation date — so anyone can reproduce the headline numbers locally without re-running retrieval.

Headline result

The Phase 2 Anthral adapter (Qwen3.5-9B-Base + LoRA r=16, GRPO on mixed binary / free-form questions) outperforms gpt-4o-mini by +5.0 accuracy points on this set, with the same prompt and retrieval:

Model	n	Acc	Brier	Reward
Anthral Phase 2 (Qwen3.5-9B-Base + LoRA)	320	60.9 %	0.259	+0.350
Anthral Phase 1 (binary-only LoRA)	320	57.5 %	0.269	+0.306
Qwen/Qwen3.5-9B-Base (untrained, same retrieval)	320	56.3 %	0.256	+0.307
gpt-4o-mini (same retrieval, same prompt)	320	55.9 %	0.278	+0.281

Reward = signed Brier (+1 − (1 − p)² if correct else −p², range [−1, 1]). Higher is better.

The full per-row scored output for each row above is in this dataset's companion artifacts on S3.

What's in the set

Each row is one yes/no forecasting question pulled from political and geopolitical event clusters in the GDELT 2.0 corpus (Aug 2025 – Apr 2026). The question is paired with the news evidence that was available strictly before the question creation date — so the model is never shown information that leaks the answer.

Schema

Column	Type	Description
qid	string	Question ID. Stable within this release.
question_title	string	The forecasting question itself.
background	string	One- or two-sentence summary of the event context.
resolution_criteria	string	How the answer is defined and what counts as a match.
answer_type	string	Always yes_no in this dataset.
ground_truth_answer	string	Either yes or no.
question_start_date	string (YYYY-MM-DD)	The cutoff. No information from this date or later is allowed when forecasting.
resolution_date	string (YYYY-MM-DD)	When the question's outcome is known.
cutoff_mode	string	Retrieval cutoff policy used (here: by question creation date).
cutoff	string (YYYY-MM-DD)	The actual cutoff date applied during retrieval.
evidence_articles	list[struct]	Retrieved news chunks. Each: {chunk_id, url, date, domain, text}. All dated strictly before the cutoff.

Quick start

from datasets import load_dataset

ds = load_dataset("rajatagarwal457/pg320-forecast-eval", split="test")
print(len(ds), "questions")
print(ds[0]["question_title"])
print("evidence chunks:", len(ds[0]["evidence_articles"]))

Reproducing the headline number with the Anthral Phase 2 adapter

from datasets import load_dataset
from openai import OpenAI

ds = load_dataset("rajatagarwal457/pg320-forecast-eval", split="test")
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")  # vLLM, see below

PROMPT = """You are a calibrated forecaster. Read the question, background, resolution criteria, and the news evidence (all dated before the question creation). Then answer.

Question: {q}
Background: {bg}
Resolution: {rc}

News evidence:
{evidence}

Reply ONLY with this exact format:
<answer>yes</answer> or <answer>no</answer>
<probability>0.XX</probability>
"""

def render(row):
    ev = "\n\n".join(f"[{c['date']} · {c['domain']}] {c['text'][:1500]}"
                     for c in row["evidence_articles"])
    return PROMPT.format(q=row["question_title"], bg=row["background"],
                         rc=row["resolution_criteria"], evidence=ev)

n_correct = 0
for row in ds:
    r = client.chat.completions.create(
        model="phase2",                     # served by vLLM with --lora-modules
        messages=[{"role": "user", "content": render(row)}],
        temperature=0.0, max_tokens=4000,
    )
    text = r.choices[0].message.content
    # parse <answer>yes/no</answer>, compare to row["ground_truth_answer"]
    ...

To serve the Phase 2 adapter:

vllm serve Qwen/Qwen3.5-9B-Base \
  --enable-lora \
  --lora-modules phase2=https://s3.eu-west-1.amazonaws.com/research.anthral.com/adapters/qwen3.5-9b-gdelt-mixed-grpo-phase2 \
  --max-loras 1 --max-lora-rank 16 --dtype bfloat16

(Adapters are at https://research.anthral.com — see that page for download links and a PEFT-only quick-load.)

Provenance & integrity

• The 320 questions are sampled from political and geopolitical event clusters in rajatagarwal457/gdelt-forecast-binary, built from the GDELT 2.0 news corpus.
• Retrieval uses TF-IDF ∪ OpenAI-embed FAISS over the same corpus, with a hard date-filter article_date < question_start_date (no leakage by construction).
• The Anthral Phase 2 adapter was trained on a disjoint training split — none of these 320 questions appear in the GRPO training data.

Citation

@misc{anthral_pg320_2026,
  author = {Anthral Labs},
  title  = {pg-320: politics/geopolitics forecasting eval},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rajatagarwal457/pg320-forecast-eval},
}

License

Released under CC-BY-4.0.